脑皮层下结构分割问题是神经科及其他相关疾病计算机辅助诊断和治疗的基础。通过分割和分析核磁共振图像中的脑结构，可以对自闭症谱系障碍、脑卒中、脑肿瘤等疾病进行早期诊断和治疗。为解决精准脑结构分割的问题，本文基于深度学习理论，提出具有继承性的两部分工作，即DenseMedic和ACNN网络的核磁共振图像脑皮层下结构的分割算法。

在第一部分的工作中，首先，OreoDown方法通过较早地增大卷积核的步长增大了特征感受野的增长速度，并使用不变尺寸的卷积层夹心式地恢复网络深度，使速度的增加带来有效的感受野增加；其次，DenseMedic使用DenseNet的构造思想实例化OreoDown框架，通过密集连接的特征提取操作来获取多尺度的上下文信息；最后，在各层中使用混合空洞卷积进一步扩大感受野，解决了特征感知过于粗糙的问题。在第二部分的工作中，首先，提出了层与层之间的接力连接，并以此将OreoDown实例化为单路径的ACNN；其次，在单路径的ACNN中央进行切割和划分，参数量没有任何增减，以多路径的ACNN神经网络在OreoDown框架下统一了单模态和多模态的分割任务；最后，在多路径的ACNN中为单模态图像引入先验信息，使灰度分布略有差异的单模态图像分别作为各个ACNN路径的输入，进一步提升了在单模态任务中的分割性能。

本文在公开的IBSR和MRBrainS18脑皮层下结构分割数据集上进行了充分的实验，以Dice相似度系数(Dice Similarity Coefficient,DSC)、交并比(Intersection over Union,IoU)、95% Hausdorff表面距离(95% Hausdorff Surface Distance,HSD95)和平均表面距离(Average Surface Distance,ASD)评价各个神经网络的分割性能。在两个数据集上同时进行的消融实验证明了本文提出的各个方法的有效性，DenseMedic在单模态T1-MR和ACNN在单模态T1-MR、多模态T1-MR+FLAIR-MR、多模态T1-MR+T1IR-MR与各个先进算法的性能对比也说明了本文的两部分工作在分割结果上的准确性和鲁棒性。实验结果表明，所提出的算法使分割出的脑结构与真实结构在区域上有更多的重叠、在轮廓上更加相似，可以更好地完成各个脑皮层下结构的分割。在临床应用中，对脑皮层下结构的精准分割将有助于准确测量相关疾病诊断的关键指标，并实现快速的计算机辅助治疗。

Subcortical structure segmentation is the basis of computer-aided diagnosis and treatment of neurology and other related diseases. By segmenting and analyzing brain structures in MRI images, early diagnosis and treatment of diseases can be performed, such as autism spectrum disorders, stroke, brain tumors and etc. In order to solve the problem of accurate subcortical segmentation, this paper proposes two parts of inherited work based on the deep learning theory, namely the DenseMedic and ACNN networks, for segmenting brain anatomies of MRI images.

In the first work of DenseMedic, firstly, the OreoDown method increases the growth rate of the characteristic receptive field by increasing the stride of convolutions in early layers, and uses convolutions with constant input and output sizes to restore the network depth in a sandwich-like manner, so that the increase in growth rate brings an effective receptive field increase; secondly, DenseMedic utilizes the construction theory of DenseNet to instantiate the OreoDown framework, while obtaining multi-scale context information through densely connected feature extracting layers; finally, the introduction of hybrid dilated convolutions in each layer further expands the receptive field, solving the possibly defect of rough feature extraction. In the second work of ACNN, firstly, the alternate connection through layers is proposed, and OreoDown is then instantiated as a single-path ACNN; secondly, the division of multi-path ACNN is performed in the center of the single-path ACNN, which unifies the single-modal and multi-modal segmentation tasks in the OreoDown framework without any changes in the parameter amount; finally, the priori information of singe-modal MR images is introduced to the multi-path ACNN to differentiate the gray distribution in separate paths, which further improves the segmentation performance in single-modal tasks.

This paper conducted sufficient experiments on the public IBSR and MRBrainS18 datasets for subcortical segmentation, utilizing Dice Similarity Coefficient (DSC), Intersection over Union (IoU), 95% Hausdorff surface distance (HSD95) and Average Surface Distance (ASD) to evaluate the segmentation performance of neural networks. Simultaneous ablation experiments on both datasets prove the effectiveness of the proposed methods. The segmenting comparisons between DenseMedic and other networks in T1-modal tasks as well as ACNN and others in T1 single-modal, T1+FLAIR bimodal and T1+T1IR bimodal tasks also proves state-of-the-art segmenting performance of DenseMedic and ACNN. The experimental results show that the segmented subcortical structures and corresponding ground truths have more overlaps in target regions and more similarity in external profiles, which indicates that DenseMedic can effectively accomplish the segmentation of major subcortical structures. In clinical applications, the presented DenseMedic will help to accurately measure the key indicators for the central nervous system related diseases and provide rapid computer-aided diagnosis and treatment.